Class Imbalance Mitigation in Hybrid LSTMCNN Architectures for Network Intrusion Detection in IoT

The rapid development in the field of IOT has introduced complex security issues due to the inherent characteristic of IOT devices being prone to vulnerabilities. Network Intrusion Detection Systems (NIDS) are crucial for securing these environments in these scenarios. The integration of Machine Learning and Deep Learning into NIDS has tremendously improved the accuracy and precision of threat detection in these systems, However, class imbalance in training datasets, where normal traffic far outweighs the attack traffic, poses a significant challenge during the development of these systems. This research proposes a hybrid Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) architecture that effectively addresses the challenge of class imbalance. The model utilizes LSTM's ability to capture temporal dependencies alongside the spatial feature extraction capabilities of CNN while implementing specialized techniques, Weighted Random Sampling, and Focal Loss Function. Experiments conducted on the TON IOT dataset resulted in the model's effectiveness, achieving 91.65% validation accuracy, 0.99 precision for normal traffic, and 0.92 precision for attack traffic. In addition, an F1-Score of 0.92 shows that the proposed approach outperforms baseline models including Logistic Regression, Random Forest, standalone CNN, LSTM-AE, and GRUCNN in threat detection. The hybrid architecture with class imbalance mitigation demonstrates a robust performance particularly for the minority attack class offering a practical solution for both know and zero-day attack detection.